Usually the systems of detection of local leakage from waste disposals are uniformly distributed as a matrix in the domain of the potential fault. In the case of leakage, the position of the source is indicated by the position of the activated sensor. A similar method is used for early warning of sudden creation of cavities beneath a building structure, due to underground industrial activity. In general, this technique allows to discover a local action by measuring some characteristic values in the close neighbourhood of the action itself. Since the position of the incident is unknown a priori, the number of sensors in this detection system is high. It could be strongly decreased - and therefore the costs could be significantly reduced - if we could identify the parameters characterising the action (leakage source position and is intensity; or position of the cavity and its diameter) from observations of the perturbation due to this incident, made far from its location. The proposed approach is based on a finite number of observations performed in a priori fixed points by varying the position of the action. In this work, virtual observations will be used, that is data are taken from numerical solutions of the problem. For example, for a steady state convection problem, assuming given concentration field values in a few measurement points and hydraulic head values in the same piezometers, the source of the concentration, its intensity and diffusivity vector are deduced using Artificial Neural Networks (ANNs) for solving a family of inverse problems. ANNs are trained with data extracted from Finite Difference (FD) solution of a classical convection problem for small Peclet number. ANN is understood here as a numerical approximation of the inverse relation, with respect to the direct solution of the differential equation for concentration. In our previous works [1], [2] we proposed an algorithm of this kind of identification. In this paper - besides advection and diffusion - adsorption, that allows to share the concentration between the solid skeleton and the fluid in the pores, is added to the model. We present some results concerning non steady processes. Furthermore, a sensitivity analysis on the accuracy of the observation of far perturbations is proposed. This is strongly connected with a definition of the so called “neutral background” that characterizes the “normal exploitation”. The model is exemplified with a case study of a design of a hydraulic barrier constructed by several injections, where the border of the impervious domain is given, and parameters of the injections are to be found as a solution of an inverse problem in the frame of the presented method.

### Identification of parameters of a local action from inverse analysis of perturbation of far fields with ANNs

#####
*BOSO, DANIELA;*

##### 2013

#### Abstract

Usually the systems of detection of local leakage from waste disposals are uniformly distributed as a matrix in the domain of the potential fault. In the case of leakage, the position of the source is indicated by the position of the activated sensor. A similar method is used for early warning of sudden creation of cavities beneath a building structure, due to underground industrial activity. In general, this technique allows to discover a local action by measuring some characteristic values in the close neighbourhood of the action itself. Since the position of the incident is unknown a priori, the number of sensors in this detection system is high. It could be strongly decreased - and therefore the costs could be significantly reduced - if we could identify the parameters characterising the action (leakage source position and is intensity; or position of the cavity and its diameter) from observations of the perturbation due to this incident, made far from its location. The proposed approach is based on a finite number of observations performed in a priori fixed points by varying the position of the action. In this work, virtual observations will be used, that is data are taken from numerical solutions of the problem. For example, for a steady state convection problem, assuming given concentration field values in a few measurement points and hydraulic head values in the same piezometers, the source of the concentration, its intensity and diffusivity vector are deduced using Artificial Neural Networks (ANNs) for solving a family of inverse problems. ANNs are trained with data extracted from Finite Difference (FD) solution of a classical convection problem for small Peclet number. ANN is understood here as a numerical approximation of the inverse relation, with respect to the direct solution of the differential equation for concentration. In our previous works [1], [2] we proposed an algorithm of this kind of identification. In this paper - besides advection and diffusion - adsorption, that allows to share the concentration between the solid skeleton and the fluid in the pores, is added to the model. We present some results concerning non steady processes. Furthermore, a sensitivity analysis on the accuracy of the observation of far perturbations is proposed. This is strongly connected with a definition of the so called “neutral background” that characterizes the “normal exploitation”. The model is exemplified with a case study of a design of a hydraulic barrier constructed by several injections, where the border of the impervious domain is given, and parameters of the injections are to be found as a solution of an inverse problem in the frame of the presented method.##### Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.